Liquid AI LFM3: a 1GB open-weight agent model with a $10M cap
Liquid AI's LFM3-1.7B fits in about 1GB at 4-bit with 256K context and strong agent scores. Its license bars commercial use above $10M revenue.
Liquid AI released LFM3, a family of six open-weight models around 2B parameters. The text model, LFM3-1.7B, fits in roughly 1 GB at 4-bit, takes a 256K-token context, and posts agent and tool-calling scores that Liquid says beat Qwen3.5-9B. That is a model you can run on a laptop, a phone or a cheap box in the back office. Read the license before you build on it, though. It is not Apache, and it has a revenue ceiling.
What actually happened
Liquid announced LFM3 on October 27 in its launch post. The details:
- The family: six models from one LFM3-1.7B base. Text, vision, audio and omni (mixed input), plus two "decision" models built from the vision and omni versions.
- The text model: 1.7B parameters, 256K context, about 1 GB at 4-bit, trained on 22T tokens. A full 128K context needs about 1.6 GB of KV cache.
- Claimed scores (Liquid's own): WebArena-Verified 38.42 against 28.94 for Qwen3.5-9B. AutomationBench 41.29 against 15.10. BFCL V4 tool use 71.39. IFEval 96.66. Liquid's own table also shows where it loses: GPQA Diamond 63.37 against 80.40, and RULER at 128K, 65.56 against 87.11.
- Where it runs: per the Hugging Face model card, GGUF for llama.cpp, Ollama and LM Studio, plus MLX, ONNX, vLLM and SGLang. 16 languages, Spanish included.
- The license: LFM Open License v1.0. Its license text says commercial use is not licensed for an entity with annual revenue of $10 million or more.
Why a 1GB agent model matters for your business
Most small-business AI tasks are narrow: read an email, pick a tool, fill a field, route the ticket. You do not need a frontier model at frontier prices for that. A 1.7B model that follows instructions well and calls tools reliably can run on hardware you already own, with zero per-call cost and no customer data leaving the building.
Two cautions. First, every number above is Liquid's own. Run your own evals on your own tickets before you trust them. Second, the $10M line matters. If you are under it today but plan to grow past it, or you are building for a client who is already over it, LFM3 is a licensing problem waiting to happen. Keep the model behind an interface you can swap, so moving to an Apache-licensed model later is a config change.
Key takeaways
- LFM3-1.7B: 256K context, about 1 GB at 4-bit, GGUF/MLX/ONNX builds on Hugging Face
- Liquid claims it beats Qwen3.5-9B on WebArena-Verified and AutomationBench, and admits it trails on GPQA and long-context RULER
- The LFM Open License v1.0 does not cover commercial use by entities with $10M+ annual revenue
- Good fit for narrow, local tool-calling tasks. Run your own evals and keep the model swappable
Small models, real work, no lock-in. We build local and hosted AI workflows behind one interface, so switching models or licenses never means a rewrite. See how we build, or ask us whether a local model fits your workflow.
Sources: Liquid AI, Hugging Face model card, LFM Open License v1.0.
- #liquid-ai
- #lfm3
- #small-language-models
- #on-device-ai
- #open-weights
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Nvidia floats $2,000-per-GPU yearly fee for agent software
Jensen Huang put Nvidia's software license at about $2,000 per GPU a year at GTC Berlin. Here's what it means for AI agent hosting costs and your vendor quotes.
Read itNVIDIA Nemotron 3.5 Bolt E4B: a 4GB local model to post-train
NVIDIA's Nemotron 3.5 Bolt E4B runs in about 4GB of VRAM with tool calling and 64K context. NVIDIA says post-train it before production. Here's why.
Read it